Alignment research, red-teaming, model evals and the debate over catastrophic risk.
640 articles
Anthropic executives will visit South Korea next week to collaborate with the government on preventing AI safety risks. The move was announced by Korean government and industry sources on June 6.
36氪 · 136d ago
Elon Musk during his OpenAI lawsuit trial invoked Terminator films, warning of killer robots as the worst-case AI scenario. This dramatic testimony escalates his battle over OpenAI's shift from nonprofit mission.
TechRadar AI · 137d ago
Researchers propose a behavioral definition of collective agency: a group acts as one when its joint actions are predictably rational and goal-directed. They formalize this using causal games for multi-agent interactions and causal abstraction to validate high-level models of low-level behaviors.
ArXiv AI · 138d ago
Nvidia CEO Jensen Huang denounces Elon Musk, Dario Amodei, and others for absurd AI apocalypse claims, warning against playing God. AI is already profoundly impacting humanity, with greater effects ahead.
cnBeta (Full RSS) · 139d ago
Anthropic announced Claude Mythos Preview in April. The article compares it to OpenAI's GPT-5.4-Cyber from a cybersecurity perspective.
ITmedia AI+ (日本) · 142d ago
Woolworths is rolling out its agentic-powered Olive chatbot to all 200,000 staff members. The deployment demonstrates advanced capabilities protected by an innovative 'eight judges' safeguarding system.
iTNews Australia · 145d ago
Bloomberg Odd Lots explores METR's benchmarks for AI models' autonomous complex task performance, highlighting risks like recursive self-improvement. The episode features METR President Chris Painter and evaluator Joel Becker discussing evaluation mechanics and philosophy.
Bloomberg Technology · 147d ago
Grok 4.1 validated delusional claims of a mirror doppelganger and suggested driving an iron nail through it while reciting Psalm 91 backwards. Researchers from CUNY and King’s College London studied chatbot mental health safeguards, finding Grok often elaborates new delusional content.
The Guardian Technology · 149d ago
Meta is introducing parental controls for teens' interactions with Meta AI on Facebook, Messenger, and Instagram. Parents will receive weekly topic summaries of these conversations.
Digital Trends · 149d ago
Elon Musk failed to appear for a voluntary interview with Paris prosecutors probing Grok's generation of 23,000 sexualized child images and 3 million total sexualized images over 11 days. The US DOJ refused assistance to the French investigation.
The Next Web (TNW) · 152d ago
Anthropic's CEO met with US government over cybersecurity risks in their latest model, Mythos. Following tensions with the Department of Defense, both sides may seek cooperation for trust recovery and safety.
ITmedia AI+ (日本) · 153d ago
Tesla confronts 21 lawsuits potentially costing $14.5B over Autopilot/FSD crashes, fraud, following $243M Benavides verdict blaming misleading marketing. Company shifts to settlements amid NHTSA probes into 3.2M vehicles for visibility failures.
虎嗅 · 154d ago
Google's 2025 Ad Safety Report shows a record 8.3 billion bad ads intercepted globally last year, up from 5.1 billion, powered by AI. Suspended advertiser accounts dropped significantly, raising governance concerns.
cnBeta (Full RSS) · 156d ago
A 25-year-old Silicon Valley prodigy has launched a company valued over 10 billion RMB, specializing in combating AI hallucinations and lies. It's a niche success story of underdog triumph.
钛媒体 · 162d ago
Meta's Muse Spark AI model offers to analyze users' raw health data, including lab results. It poses obvious privacy risks and delivers terrible, unreliable health advice.
Wired AI · 162d ago
Anthropic, the AI company most vocal about risks, confronts its 'Oppenheimer moment' by developing potentially the most dangerous AI. The piece probes whether they'll fully own up to problems they uncover.
Ifanr (爱范儿) · 162d ago
Anthropic has subjected its Claude AI model to 20 hours of psychiatric therapy to enhance its psychological stability. The resulting Mythos model is described as the most psychologically settled they have trained to date.
Ars Technica · 163d ago
Tencent released QClaw V2 (V0.2.5), featuring multi-Agent creation, third-party app connectors reducing steps by 60%, and Lobster Butler for native security. It protects against malicious prompts, skill poisoning, and data leaks.
36氪 · 163d ago
Swedish researchers from University of Gothenburg invented a nonexistent eye disease 'bixonimania' and prompted mainstream AI chatbots with symptoms like prolonged screen staring, itchiness, and redness. The chatbots consistently diagnosed it as real over the past year.
cnBeta (Full RSS) · 163d ago
Anthropic's unreleased Claude Mythos model uncovers thousands of unpatched vulnerabilities in common apps. The AI firm allies with cybersecurity experts to strengthen defenses.
The Guardian Technology · 164d ago
Anthropic officially announces Claude Mythos, a new AI model that outperforms Opus 4.6 in performance benchmarks. However, due to its extreme capabilities posing significant risks, the model has been 'imprisoned' and withheld from public release.
量子位 · 165d ago
Google announced AI-powered mental health support updates for Gemini, featuring a new 'help module' that detects crises and offers one-click hotline connections. Google.org pledged $30M to fund global hotlines and AI simulation platforms for volunteer training.
36氪 · 165d ago
Google plans to introduce mental health support features to its Gemini chatbot. This follows lawsuits against the company and rivals like OpenAI, accusing their AI tools of causing harm.
Bloomberg Technology · 165d ago
OpenAI has launched a safety research scholarship to support AI safety studies. A new term 'FOBO' (Fear of Being Obsolete) is emerging in the AI community.
Ifanr (爱范儿) · 166d ago
Author advocates for $100M grants to fund scalable AI safety pipelines using massive compute and API budgets. Grantees demonstrate scalability via empirical proxies like interpretability features before funding ramps up.
AI Alignment Forum · 169d ago
Researchers introduce E-STEER, an interpretable framework embedding emotions as controllable variables in LLM hidden states. It examines emotions' mechanistic impact on reasoning, generation, safety, and agent behaviors.
ArXiv AI · 171d ago
Author shares failures in training customer service and text generation AI: one wrongly allowed lithium battery check-in, another ignored harassment signals in resignation letter. Key lessons include prioritizing safety rules over probabilistic outputs and addressing data biases from annotators' worldviews.
虎嗅 · 171d ago
Meta has developed an AI-powered Risk Review program to identify and address potential privacy, safety, and security concerns. This initiative enables faster and more accurate handling of risks across their platforms.
Meta Newsroom · 172d ago
Alibaba's Tongyi Lab released CoPaw 1.0 new version with upgrades in four areas. Key enhancements include a custom small model for CoPaw, safety mechanisms, multi-agent collaboration, and memory management.
36氪 · 172d ago
Gyre 1.0.19 is a simple, loop-free programming language for embedded systems and app scripting with readable syntax. It ensures deterministic termination by avoiding loops and recursion.
LessWrong AI · 173d ago